The style transfer paper is very unique in the ranks of seminal deep learning papers. The authors are more active in neuroscience than in machine learning, but showed the amazing results you can get when you apply knowledge about the human brain to experiments on neural networks. The paper is also unique in its readability: it contains one “Methods” section which explains the style transfer loss function, but outside of that section, the paper is readable for almost anyone that knows a thing or two about neural networks.
The chief contribution of this work is the fact that the style and content of an image are separable. Even after they lay out the theoretical foundation, this is the kind of thing where you really have to see it to believe it. We talked about why this makes sense intuitively, how “style” and “content” vary across deeper and deeper layers of the VGG CNN, and then we looked at examples of both (1) isolating the content/style of an image and visualizing through gradient descent and (2) combining the style of image A with the content of image B. We think that these works of art are some of the most fascinating and beautiful results ever produced in deep learning.